Support mode 'nearest-exact' for torch.nn.functional.interpolate

[lianakoleva]

First step towards addressing #1941, referencing ATen source.

[mruberry]

Hey @lianakoleva! Cool stuff. I made a few small comments. @beverlylytle, would you also like to take a look?

To test the implementation, we'll need to extend the "sample inputs" of interpolate, which are defined here:

lightning-thunder/thunder/tests/opinfos.py

Line 9320 in 4acf819

def interpolate_sample_generator(op, device, dtype, requires_grad, **kwargs):

In particular, I think this part:

lightning-thunder/thunder/tests/opinfos.py

Line 9363 in 4acf819

yield SampleInput(make(a_shape), size=size)

could be edited to look like:

yield SampleInput(make(a_shape), size=size)
yield SampleInput(make(a_shape), size=size, mode='nearest-exact')

So that both modes are tested. Please let us know if you have any questions about this, as the testing can be confusing at first!